Early warning identification method and system for ultra-high voltage transformer substation

Abstract

The invention discloses an early warning identification method and system for an ultra-high voltage transformer substation, and the method comprises the steps: obtaining a transformer substation image, inputting the transformer substation image into a first network early warning model for alarm prediction, and obtaining a first probability value; judging whether the first probability value is within a preset range or not, and obtaining a first judgment result; when the first judgment result indicates that the first probability value is within a preset range, inputting the transformer substation image to a second network early warning model for alarm prediction, and obtaining a second probability value; judging whether the second probability value is within a preset range or not, and obtaining a second judgment result; and when the second judgment result indicates that the second probability value is not within the preset range, refusing to identify the alarm. The method can effectively solve the contradictory problem of alarm information redundancy and alarm omission in an intelligent inspection system, can give an alarm which does not need to be alarmed and is determined to be an alarm of a certain type or is rejected to be recognized, and can improve the early warning precision of the ultrahigh-voltage transformer substation.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power systems, and more specifically, to an early warning identification method and system for an ultra-high voltage substation. Background Art

[0002] With the rise of artificial intelligence, image recognition and classification technologies are maturing and being applied across various industries. Ultra-high voltage substations, as core hubs of the power system, are characterized by large equipment volumes and complex operating environments. Traditional manual inspections are plagued by low efficiency, high risk, and incomplete data collection. While the introduction of intelligent inspection systems, using technologies like drones, robots, and multiple sensors, enables real-time monitoring of equipment status, the massive amount of data and algorithmic limitations lead to alarm overload and the risk of missed alerts.

[0003] The application of convolutional neural networks (CNNs) in the intelligent inspection system for ultra-high voltage substations begins with the collection and preprocessing of multi-source equipment image data. A training set is constructed by collecting infrared thermal images, visible light images, and abnormal status labels. Image cropping, normalization, and denoising are performed to eliminate electromagnetic interference and illumination variations. Data augmentation techniques such as rotation, flipping, and noise addition are used to address the small sample size issue. A generative adversarial network is also used to simulate complex scenarios such as rain, fog, and oil pollution to enhance data diversity. The model construction phase improves the network structure based on pre-trained models such as ResNet and the YOLO series. The CSPDarknet53 backbone network is used to extract local features such as insulator cracks and metal corrosion. The SPP module is combined to enhance multi-scale feature fusion capabilities. Parameters are fine-tuned using transfer learning on a dedicated substation dataset. Iterative training is driven by a cross-entropy loss function and the Adam optimizer. An attention mechanism is introduced to dynamically weight the correlation between infrared and visible light features. Batch normalization and dropout techniques are used to prevent overfitting and enhance the model's adaptability to equipment angle changes and partial occlusion. During the test and verification phase, the F1 score and average precision indicators are used to evaluate performance. The confusion matrix is ​​used to analyze the types of false detections and optimize the model. Pruning and quantization techniques are used to compress the parameters to address the computing power limitations of the edge. Lightweight networks are deployed to inspection robots or drones to achieve real-time detection of millimeter-level defects such as transformer bushing liquid level and lightning arrester counters. At the same time, a dual-engine cross-validation mechanism is combined to trigger multimodal fusion review when the confidence level is insufficient to reduce the risk of missed detection.

[0004] The application of dynamic threshold adjustment technology in substation intelligent inspection systems relies on a multi-stage data-driven approach, coordinated by intelligent algorithms. First, the system collects multi-dimensional parameters such as current, voltage, temperature, and vibration in real time from sensors deployed on key equipment such as transformers, circuit breakers, and transformers. A regularization module cleans and normalizes the raw data to eliminate dimensional discrepancies and noise. Subsequently, dimensionality reduction algorithms such as principal component analysis are used to extract core features. Clustering models are then used to classify the data and identify data distribution patterns under different operating conditions. For example, weighted cosine similarity is used to demarcate the boundaries between normal and abnormal data. During the dynamic threshold generation phase, the system integrates a time series prediction model to analyze equipment operating trends. A sliding window is used to calculate the mean and standard deviation of real-time data, and dynamic threshold intervals are calculated using the quartile method. An adaptive algorithm is also implemented to adjust threshold sensitivity based on load fluctuations or environmental factors. During anomaly detection, the system compares the scores of real-time data with predicted values. If the threshold value exceeds the threshold, an alarm is triggered. Local processing is performed using edge computing nodes to reduce communication latency. At the optimization and iteration level, the system continuously monitors device aging or data drift through a feedback mechanism. It then optimizes model parameters using reinforcement learning algorithms and incorporates expert experience to fine-tune thresholds, forming a closed-loop optimization process known as "data-model-knowledge." Ultimately, periodic recalibration ensures that thresholds dynamically match the actual device state, significantly improving anomaly identification accuracy and system robustness.

[0005] The intelligent inspection and maintenance systems currently deployed in ultra-high voltage substations utilize advanced AI-based technologies such as image recognition and pattern analysis for alarm monitoring. However, significant challenges exist in practical applications: if the system's alarm threshold is set too high, while it can effectively detect equipment anomalies, the massive amount of alarm information not only increases the workload of operators, but can also induce "alarm fatigue," leading to delayed responses to critical alarms or even their being ignored. Conversely, if the alarm threshold is set too low, while it can reduce invalid alarms, it can also miss hidden fault characteristics, posing a significant safety hazard.

[0006] Therefore, a method for early warning identification of ultra-high voltage substations is needed. Summary of the Invention

[0007] The present invention proposes a method and system for early warning identification of an ultra-high voltage substation to solve the problem of how to improve the early warning accuracy of the ultra-high voltage substation.

[0008] In order to solve the above problem, according to one aspect of the present invention, a method for early warning identification of an ultra-high voltage substation is provided, the method comprising:

[0009] Obtaining a substation image, and inputting the substation image into a first network early warning model for alarm prediction to obtain a first probability value;

[0010] Determine whether the first probability value is within a preset range, and obtain a first determination result;

[0011] When the first judgment result indicates that the first probability value is within a preset range, inputting the substation image into a second network early warning model for alarm prediction to obtain a second probability value;

[0012] Determine whether the second probability value is within a preset range, and obtain a second determination result;

[0013] When the second judgment result indicates that the second probability value is not within a preset range, the alarm is rejected.

[0014] Preferably, the method further comprises:

[0015] When the first judgment result indicates that the first probability value is not within the preset range, determine whether the first probability value is less than a first preset threshold; if the first probability value is less than the first preset threshold, reject the identification alarm; if the first probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0016] Preferably, the method further comprises:

[0017] When the second judgment result indicates that the second probability value is not within the preset range, determine whether the second probability value is less than the first preset threshold; if the second probability value is less than the first preset threshold, reject the identification alarm; if the second probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0018] Preferably, the method further comprises:

[0019] Constructing the first network early warning model and performing model training on the first network early warning model; wherein the first network early warning model is constructed based on a discrete Hopfield network;

[0020] Construct the second network early warning model and perform model training on the second network early warning model; wherein, the second network early warning model is constructed based on the interpretable convolutional neural network ECNN.

[0021] Preferably, during the training process of the first network warning model, the training sample images are binarized, the outer product of the training sample images and the weight matrix are calculated, the target pattern is stored, and a discrete Hopfield network is constructed. The number of neurons n is defined to be equal to the number of pixels in the image, the weight matrix W and the threshold Ti are initialized, and the energy function is designed to ensure that the energy monotonically decreases during the dynamic evolution of the network and converges to a stable state.

[0022] Preferably, during the training process of the second network warning model, the training sample image is first resized, standardized and data enhanced, and then the convolution layer parameters are retained, the classification layer after GAP is trained, and then the feature maps of the n channels output by the last convolution layer are weighted and summed and mapped to the original image to obtain the importance of different areas.

[0023] Preferably, when using the first network early warning model for alarm prediction, a noisy input X0 is generated by randomly flipping some pixels, and it is used as the initial state. Then, neurons i are randomly selected one by one, and their states are updated according to the neighbor states and weights. During the update process, the energy decreases monotonically. When the energy reaches a local minimum, the network converges to a stable state. By comparing the stable state with the training pattern, the pattern with the smallest Euclidean distance is selected as the classification result.

[0024] Preferably, when using the second network early warning model for alarm prediction, the weight after GAP is directly obtained after inputting data. And the last layer of convolutional feature map f k , based on the weight after GAP And the last layer of convolutional feature map f k Map the heat map, normalize the heat map, scale the heat map to the range of 0 to 1, and then use bilinear interpolation to enlarge the heat map to the original image size. Superimpose the heat map on the original image to show the feature areas that the specified category focuses on on the original image.

[0025] According to another aspect of the present invention, there is provided an early warning identification system for an ultra-high voltage substation, the system comprising:

[0026] a first prediction unit, configured to obtain a substation image, and input the substation image into a first network early warning model for alarm prediction to obtain a first probability value;

[0027] a first judgment result obtaining unit, configured to determine whether the first probability value is within a preset range, and obtain a first judgment result;

[0028] a second prediction unit, configured to input the substation image into a second network early warning model for alarm prediction to obtain a second probability value when the first judgment result indicates that the first probability value is within a preset range;

[0029] a second determination result obtaining unit, configured to determine whether the second probability value is within a preset range, and obtain a second determination result;

[0030] The identification alarm rejection unit is configured to reject the identification alarm when the second judgment result indicates that the second probability value is not within a preset range.

[0031] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, any step of a method for early warning identification of an ultra-high voltage substation is implemented.

[0032] According to another aspect of the present invention, the present invention provides an electronic device, including:

[0033] The computer-readable storage medium described above; and

[0034] One or more processors are configured to execute the program in the computer-readable storage medium.

[0035] The present invention provides an ultra-high voltage substation early warning identification method and system, including: obtaining a substation image, and inputting the substation image into a first network early warning model for alarm prediction to obtain a first probability value; judging whether the first probability value is within a preset range to obtain a first judgment result; when the first judgment result indicates that the first probability value is within the preset range, inputting the substation image into a second network early warning model for alarm prediction to obtain a second probability value; judging whether the second probability value is within the preset range to obtain a second judgment result; when the second judgment result indicates that the second probability value is not within the preset range, refusing to identify the alarm. The method of the present invention achieves a method of refusing to identify different situations by designing two networks to coordinate with each other, which can effectively solve the contradiction between alarm information redundancy and omission in traditional intelligent inspection systems. After the data is input, it can give an alarm that does not really need to be alerted, an alarm that is determined to be a certain type of alarm, or an alarm that refuses to be identified, thereby improving the accuracy of ultra-high voltage substation early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] A more complete understanding of exemplary embodiments of the present invention may be obtained by referring to the following drawings:

[0037] Figure 1 Flowchart of a method 100 for early warning identification of an ultra-high voltage substation according to an embodiment of the present invention;

[0038] Figure 2 1. It is an overall flow chart of the early warning identification process of the ultra-high voltage substation according to the embodiment of the present invention;

[0039] Figure 3 2. A schematic structural diagram of a discrete Hopfield network according to an embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the structure of an interpretable convolutional neural network according to an embodiment of the present invention;

[0041] Figure 5 FIG. 5 is a structural diagram of an early warning identification system 500 for an ultra-high voltage substation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to provide a thorough and complete disclosure of the present invention and to fully convey the scope of the present invention to those skilled in the art. The terminology used in the exemplary embodiments shown in the accompanying drawings is not intended to limit the present invention. In the accompanying drawings, identical elements are denoted by the same reference numerals.

[0043] Unless otherwise specified, the terms used herein (including technical terms) have the meanings commonly understood by those skilled in the art. In addition, it is understood that terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

[0044] In order to solve the contradiction between alarm information redundancy and omission in traditional intelligent inspection systems, the present invention proposes an early warning identification method for ultra-high voltage substations. Based on the improved DHNN (discrete Hopfield network) and ECNN (interpretable convolutional neural network), prediction is performed to give alarms that do not really need to be warned, are determined to be a certain type of alarm, or refuse to be identified, which can improve the early warning accuracy of ultra-high voltage substations.

[0045] Figure 1 FIG. 1 is a flow chart of an early warning identification method 100 for an ultra-high voltage substation according to an embodiment of the present invention. Figure 1As shown, the early warning identification method 100 for an ultra-high voltage substation provided by an embodiment of the present invention, by designing two networks to coordinate with each other, can provide a method for rejecting identification for different situations. It can effectively solve the contradiction between redundant alarm information and missed reports in traditional intelligent inspection systems. After inputting data, it can provide alarms that are indeed unnecessary, determine whether they are certain types of alarms, or reject identification, which can improve the early warning accuracy of ultra-high voltage substations. The early warning identification method 100 for an ultra-high voltage substation provided by an embodiment of the present invention begins at step 101. In step 101, a substation image is obtained and the substation image is input into a first network early warning model for alarm prediction to obtain a first probability value.

[0046] In step 102, it is determined whether the first probability value is within a preset range, and a first determination result is obtained.

[0047] In step 103, when the first judgment result indicates that the first probability value is within a preset range, the substation image is input into a second network early warning model for alarm prediction to obtain a second probability value.

[0048] In step 104, it is determined whether the second probability value is within a preset range, and a second determination result is obtained.

[0049] In step 105, when the second judgment result indicates that the second probability value is not within a preset range, the alarm is rejected.

[0050] Preferably, the method further comprises:

[0051] When the first judgment result indicates that the first probability value is not within the preset range, determine whether the first probability value is less than a first preset threshold; if the first probability value is less than the first preset threshold, reject the identification alarm; if the first probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0052] Preferably, the method further comprises:

[0053] When the second judgment result indicates that the second probability value is not within the preset range, determine whether the second probability value is less than the first preset threshold; if the second probability value is less than the first preset threshold, reject the identification alarm; if the second probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0054] Combine Figure 2As shown, in the present invention, first, a substation image is acquired. Then, a first network prediction model constructed and trained based on a first network is used to perform alarm prediction and obtain a first probability value. If the obtained first probability value is greater than a set value T2, an alarm is determined; if the obtained probability value is less than a set value T1, no alarm is determined. If the first probability value is within the set interval [T1, T2], a second network prediction model constructed and trained based on a second network is used to perform prediction and obtain a second probability value. If the second probability value is greater than a set value T2, an alarm is determined; if the obtained probability value is less than a set value T1, no alarm is determined; if the second probability value is still within the set interval [T1, T2], a recognition rejection alarm is issued. The substation image includes images such as the equipment appearance, meter readings, and infrared images of the equipment body and connection parts.

[0055] Preferably, the method further comprises:

[0056] Constructing the first network early warning model and performing model training on the first network early warning model; wherein the first network early warning model is constructed based on a discrete Hopfield network;

[0057] Construct the second network early warning model and perform model training on the second network early warning model; wherein, the second network early warning model is constructed based on the interpretable convolutional neural network ECNN.

[0058] Preferably, during the training process of the first network warning model, the training sample images are binarized, the outer product of the training sample images and the weight matrix are calculated, the target pattern is stored, and a discrete Hopfield network is constructed. The number of neurons n is defined to be equal to the number of pixels in the image, the weight matrix W and the threshold Ti are initialized, and the energy function is designed to ensure that the energy monotonically decreases during the dynamic evolution of the network and converges to a stable state.

[0059] Preferably, during the training process of the second network warning model, the training sample image is first resized, standardized and data enhanced, and then the convolution layer parameters are retained, the classification layer after GAP is trained, and then the feature maps of the n channels output by the last convolution layer are weighted and summed and mapped to the original image to obtain the importance of different areas.

[0060] Preferably, when using the first network early warning model for alarm prediction, a noisy input X0 is generated by randomly flipping some pixels, and it is used as the initial state. Then, neurons i are randomly selected one by one, and their states are updated according to the neighbor states and weights. During the update process, the energy decreases monotonically. When the energy reaches a local minimum, the network converges to a stable state. By comparing the stable state with the training pattern, the pattern with the smallest Euclidean distance is selected as the classification result.

[0061] Preferably, when using the second network early warning model for alarm prediction, the weight after GAP is directly obtained after inputting data. And the last layer of convolutional feature map f k , based on the weight after GAP And the last layer of convolutional feature map f k Map the heat map, normalize the heat map, scale the heat map to the range of 0 to 1, and then use bilinear interpolation to enlarge the heat map to the original image size. Superimpose the heat map on the original image to show the feature areas that the specified category focuses on on the original image.

[0062] In the present invention, the first network prediction model is constructed based on the discrete Hopfield network DHNN, and the second network prediction model is constructed based on the interpretable convolutional neural network ECNN.

[0063] DHNN is a single-layer fully connected feedback neural network, and its network structure is shown in the figure below. Figure 3 As shown, where Ti is the threshold, S i The operation process of DHNN is as follows: when an input is given to the network, the network will change its state according to a certain evolution law until the state of the network no longer changes, and it reaches a stable state.

[0064] In the present invention, the training phase of the first network prediction model includes:

[0065] 1. Data preprocessing and feature encoding

[0066] First, the image is binarized. As shown in Equation (1), the grayscale values ​​of the image pixels are normalized into a binary state, where 1 represents activation and -1 represents inhibition. To reduce the amount of computation, the high-resolution image is downsampled to a low resolution.

[0067]

[0068] 2. Weight matrix calculation

[0069] The weight matrix is ​​calculated by summing the outer products of the training samples and storing the target modes, including warning, no warning and rejection of warning. The weight matrix is ​​calculated as shown in formula (2).

[0070]

[0071] Among them, V μ is the μth training sample, W ii =0, the weight matrix satisfies W ij =W ji .

[0072] 3. Network initialization

[0073] Construct a discrete Hopfield network, define the number of neurons n equal to the number of image pixels, initialize the weight matrix W and threshold Ti, and design an energy function to ensure that the energy decreases monotonically during the dynamic evolution of the network and converges to a stable state.

[0074] In the present invention, the prediction stage of the first network prediction model includes:

[0075] 1. Input noise pattern

[0076] The noisy input X0 is generated by randomly flipping some pixels and used as the initial state S0=X0.

[0077] 2. Asynchronous state updates

[0078] Neuron i is randomly selected one by one, and its state is updated according to the neighbor state and weight as shown in formula (3).

[0079]

[0080] Wherein, if x≥0, sign(x)=1, otherwise it is -1.

[0081] 3. Energy function convergence

[0082] The energy function can measure the stability of the network state. The energy decreases monotonically during the update process. Its expression is shown in formula (4). When the energy reaches a local minimum, the network converges to a stable state.

[0083]

[0084] 4. Output result matching

[0085] The stable state is compared with the training pattern, and the pattern with the smallest Euclidean distance is selected as the classification result.

[0086] Compared with traditional convolutional neural networks (CNN), ECNN uses global average pooling (GAP) instead of the fully connected layer to directly compress the spatial dimension of the feature map of the last convolutional layer, retaining the semantic information of the channel dimension, thereby retaining the model's positioning feature capability. Its structure is as follows: Figure 4As shown in Figure 2. GAP usually plays a regularization role to prevent overfitting during training. The global pooling method is to expand the pooling range to the size of the entire feature map. Figure 3 As shown in the Class Activation Map (CAM), the last convolutional layer is followed by GAP, and then the Softmax layer is used for classification. The number of channels in the last convolutional layer in the figure is n, so the vector dimension obtained after GAP is n, corresponding to each channel, w1,...,w in the figure n Refers to the weight of the Softmax layer, which corresponds to the weight of a class.

[0087] In the present invention, the training phase of the second network prediction model includes:

[0088] 1. Data Processing

[0089] First, the image is preprocessed, the size is adjusted to adapt to the network structure, normalized, and data enhancement is performed to improve the generalization ability of the model.

[0090] 2. Feature Extraction

[0091] Keep the convolution layer parameters and only train the classification layer after GAP.

[0092] 3. Calculate importance

[0093] The feature map output by the last convolutional layer contains n channels. The weight of the class is used to weight the n channels of the feature map and then map it to the original image to get the importance of different regions. Here, because the feature map and the original image have different sizes, the mapping is required. CAM directly scales the feature map to the same size as the original image. Figure 1 For example, for an image, use formula (5) to calculate the score of its belonging to category 1.

[0094]

[0095] Among them, k represents the channel of the last convolutional layer, xy represents the coordinates of the feature map, and f k Represents the last layer of convolutional feature map, represents the weight of category 1 on the kth channel.

[0096] For category 1, we can obtain a map of the importance of each region of the image, as shown in formula (6), which is obtained by weighted summation of the feature maps.

[0097]

[0098] In the present invention, the prediction stage of the second network prediction model includes:

[0099] 1. Feature map and weight extraction

[0100] After inputting the data, directly obtain the weight after GAP And the last layer of convolutional feature map f k .

[0101] 2. Heatmap calculation

[0102] Normalize the heatmap, scale it to the range of 0 to 1, and then use bilinear interpolation to enlarge the heatmap to the original image size.

[0103] 3. Visualization

[0104] Overlay the heat map onto the original image to show the feature areas that the specified category focuses on on the original image.

[0105] Figure 5 FIG. 5 is a structural diagram of an early warning identification system 500 for an ultra-high voltage substation according to an embodiment of the present invention. Figure 5 As shown, the early warning identification system 500 for the ultra-high voltage substation provided by the embodiment of the present invention includes: a first prediction unit 501, a first judgment result acquisition unit 502, a second prediction unit 503, a second judgment result acquisition unit 504 and an identification alarm rejection unit 505.

[0106] Preferably, the first prediction unit 501 is used to obtain a substation image, and input the substation image into a first network early warning model to perform alarm prediction and obtain a first probability value.

[0107] Preferably, the first judgment result obtaining unit 502 is configured to determine whether the first probability value is within a preset range, and obtain a first judgment result.

[0108] Preferably, the second prediction unit 503 is configured to input the substation image into a second network early warning model for alarm prediction to obtain a second probability value when the first judgment result indicates that the first probability value is within a preset range.

[0109] Preferably, the second judgment result obtaining unit 504 is configured to determine whether the second probability value is within a preset range, and obtain a second judgment result.

[0110] Preferably, the recognition alarm rejection unit 505 is configured to reject the recognition alarm when the second judgment result indicates that the second probability value is not within a preset range.

[0111] Preferably, the system further comprises:

[0112] When the first judgment result indicates that the first probability value is not within the preset range, determine whether the first probability value is less than a first preset threshold; if the first probability value is less than the first preset threshold, reject the identification alarm; if the first probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0113] Preferably, the system further comprises:

[0114] When the second judgment result indicates that the second probability value is not within the preset range, determine whether the second probability value is less than the first preset threshold; if the second probability value is less than the first preset threshold, reject the identification alarm; if the second probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

[0115] Preferably, the system further comprises: a model building unit, configured to:

[0116] Constructing the first network early warning model and performing model training on the first network early warning model; wherein the first network early warning model is constructed based on a discrete Hopfield network;

[0117] Construct the second network early warning model and perform model training on the second network early warning model; wherein, the second network early warning model is constructed based on the interpretable convolutional neural network ECNN.

[0118] Preferably, the model construction unit, during the training process of the first network early warning model, binarizes the training sample images, calculates the outer product of the training sample images and calculates the weight matrix, stores the target pattern, and constructs a discrete Hopfield network, defines the number of neurons n equal to the number of pixels in the image, initializes the weight matrix W and the threshold Ti, and designs the energy function to ensure that the energy monotonically decreases during the dynamic evolution of the network and converges to a stable state.

[0119] Preferably, the model building unit, during the training process of the second network warning model, first resizes, standardizes and performs data enhancement processing on the training sample image, then retains the convolution layer parameters, trains the classification layer after GAP, and then performs weighted summation on the feature maps of the n channels output by the last convolution layer, and maps them to the original image, thereby obtaining the importance of different areas.

[0120] Preferably, the first prediction unit 501, when using the first network early warning model to perform alarm prediction, generates a noisy input X0 by randomly flipping some pixels, and uses it as the initial state, and then randomly selects neurons i one by one, updates its state according to the neighbor state and weight, and the energy decreases monotonically during the update process. When the energy reaches the local minimum, the network converges to a stable state, and by comparing the stable state with the training pattern, the pattern with the smallest Euclidean distance is selected as the classification result.

[0121] Preferably, the second prediction unit 503, when using the second network early warning model to perform alarm prediction, directly obtains the weight after GAP after inputting data And the last layer of convolutional feature map f k , based on the weight after GAP And the last layer of convolutional feature map f k Map the heat map, normalize the heat map, scale the heat map to the range of 0 to 1, and then use bilinear interpolation to enlarge the heat map to the original image size. Superimpose the heat map on the original image to show the feature areas that the specified category focuses on on the original image.

[0122] The early warning identification system 500 for an ultra-high voltage substation according to an embodiment of the present invention corresponds to the early warning identification method 100 for an ultra-high voltage substation according to another embodiment of the present invention, and will not be described in detail herein.

[0123] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, any step of a method for early warning identification of an ultra-high voltage substation is implemented.

[0124] According to another aspect of the present invention, the present invention provides an electronic device, including:

[0125] The computer-readable storage medium described above; and

[0126] One or more processors are configured to execute the program in the computer-readable storage medium.

[0127] The present invention has been described with reference to a few embodiments. However, it is apparent to those skilled in the art that other embodiments than the ones disclosed above are equally within the scope of the present invention.

[0128] Generally, all terms used in this disclosure are to be interpreted according to their ordinary meaning in the art, unless explicitly defined otherwise herein. All references to "a / the / the [device, component, etc.]" are to be interpreted openly as referring to at least one instance of the device, component, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not necessarily need to be performed in the exact order disclosed, unless explicitly stated otherwise.

[0129] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0132] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for early warning identification of an ultra-high voltage substation, characterized in that: The method comprises: Obtaining a substation image, and inputting the substation image into a first network early warning model for alarm prediction to obtain a first probability value; Determine whether the first probability value is within a preset range, and obtain a first determination result; When the first judgment result indicates that the first probability value is within a preset range, inputting the substation image into a second network early warning model for alarm prediction to obtain a second probability value; Determine whether the second probability value is within a preset range, and obtain a second determination result; When the second judgment result indicates that the second probability value is not within a preset range, the alarm is rejected.

2. The method according to claim 1, characterized in that The method further comprises: When the first judgment result indicates that the first probability value is not within the preset range, determine whether the first probability value is less than a first preset threshold; if the first probability value is less than the first preset threshold, reject the identification alarm; if the first probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

3. The method according to claim 1, characterized in that The method further comprises: When the second judgment result indicates that the second probability value is not within the preset range, determine whether the second probability value is less than the first preset threshold; if the second probability value is less than the first preset threshold, reject the identification alarm; if the second probability threshold is greater than the second preset threshold, issue an identification alarm; the preset range is [T1, T2], T1 is the first preset threshold, and T2 is the second preset threshold.

4. The method according to claim 1, wherein The method further comprises: Constructing the first network early warning model and performing model training on the first network early warning model; wherein the first network early warning model is constructed based on a discrete Hopfield network; Construct the second network early warning model and perform model training on the second network early warning model; wherein, the second network early warning model is constructed based on the interpretable convolutional neural network ECNN.

5. The method according to claim 4, characterized in that During the training process of the first network warning model, the training sample images are binarized, the outer product of the training sample images and the weight matrix are calculated, the target pattern is stored, and a discrete Hopfield network is constructed. The number of neurons n is defined as equal to the number of pixels in the image, the weight matrix W and the threshold Ti are initialized, and the energy function is designed to ensure that the energy decreases monotonically during the dynamic evolution of the network and converges to a stable state.

6. The method according to claim 4, characterized in that During the training process of the second network warning model, the training sample images are first resized, standardized, and data augmented. Then, the convolutional layer parameters are retained, and the classification layer after GAP is trained. Finally, the feature maps of the n channels output by the last convolution layer are weighted summed and mapped to the original image to obtain the importance of different areas.

7. The method according to claim 5, characterized in that When using the first network warning model for alarm prediction, a noisy input X0 is generated by randomly flipping some pixels and using it as the initial state. Neuron i is then randomly selected one by one and its state is updated according to the neighbor state and weight. During the update process, the energy decreases monotonically. When the energy reaches the local minimum, the network converges to a stable state. By comparing the stable state with the training pattern, the pattern with the smallest Euclidean distance is selected as the classification result.

8. The method according to claim 6, characterized in that When using the second network early warning model for alarm prediction, directly obtain the weight after GAP after inputting data And the last layer of convolutional feature map f k , based on the weight after GAP And the last layer of convolutional feature map f k Map the heat map, normalize the heat map, scale the heat map to the range of 0 to 1, and then use bilinear interpolation to enlarge the heat map to the original image size. Superimpose the heat map on the original image to show the feature areas that the specified category focuses on on the original image.

9. An early warning and identification system for an ultra-high voltage substation, characterized in that: The system comprises: a first prediction unit, configured to obtain a substation image, and input the substation image into a first network early warning model for alarm prediction to obtain a first probability value; a first judgment result obtaining unit, configured to determine whether the first probability value is within a preset range, and obtain a first judgment result; a second prediction unit, configured to input the substation image into a second network early warning model for alarm prediction to obtain a second probability value when the first judgment result indicates that the first probability value is within a preset range; a second determination result obtaining unit, configured to determine whether the second probability value is within a preset range, and obtain a second determination result; The identification alarm rejection unit is configured to reject the identification alarm when the second judgment result indicates that the second probability value is not within a preset range.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

11. An electronic device, characterized in that: include: The computer-readable storage medium of claim 10; as well as One or more processors are configured to execute the program in the computer-readable storage medium.

Images